CN102521366B - Image retrieval method integrating classification with hash partitioning and image retrieval system utilizing same - Google Patents
Image retrieval method integrating classification with hash partitioning and image retrieval system utilizing same Download PDFInfo
- Publication number
- CN102521366B CN102521366B CN 201110423143 CN201110423143A CN102521366B CN 102521366 B CN102521366 B CN 102521366B CN 201110423143 CN201110423143 CN 201110423143 CN 201110423143 A CN201110423143 A CN 201110423143A CN 102521366 B CN102521366 B CN 102521366B
- Authority
- CN
- China
- Prior art keywords
- picture
- module
- classification
- checked
- feature
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Fee Related
Links
Images
Landscapes
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses an image retrieval system integrating classification with hash partitioning, which comprises a downloading module, a classification module training module, an image classification module, a characteristic extraction module, a recording table building module, a partitioning module, a request processing module, a retrieval module, a similarity acquiring module and a result returning module. The downloading module is used for downloading images so as to build an image library, the classification model training module classifies the images in the image library according to the shapes and selects representative sample images from the image library to form a sample library firstly, then extracts characteristic descriptors of the classification bottom layer of all images in the sample library and performs trainings on the characteristic descriptors of the classification bottom layer by a support vector machine so as to obtain discriminant of each classification. Classification models are formed according to the discriminants of all the classifications. Precision ratio of the image retrieval system is increased, the problem of low recall ratio during classification mistakes is overcome, and the retrieval speed of the image retrieval system is increased integrally.
Description
Technical field
The present invention relates to the vertical searching field of content-based image, more particularly, the present invention relates to image search method and the image indexing system of a kind of integrated classification and global index.
Background technology
Existing CBIR, main retrieval mode has, based on the retrieval of classification, based on the retrieval of cluster and retrieval based on global index.Retrieval based on classification is in advance the picture classification in database, at first obtains the classification of inquiry picture during retrieval, and then retrieves similar picture in classification; Retrieval based on cluster is that all picture feature are carried out cluster, forms cluster centre, and during retrieval, at first picture to be checked searches nearest cluster centre, then searches similar picture in picture set corresponding to this cluster centre; Retrieval based on global index is that all picture feature are set up index, and picture to be checked is searched the set of similar pictures on index, then return to picture similar in set.
Yet there is following problem in existing CBIR method: when adopting the cluster mode, lost the precision of proper vector due to cluster and index, caused inquiring about accuracy rate low; When adopting mode classification, if during picture classification mistake to be checked, precision ratio and recall ratio all can reduce greatly; When adopting indexed mode, the index of setting up on the Characteristic of Image vector, its inquiry velocity is slow, can cause the retrieval of system consuming time.
Summary of the invention
The object of the present invention is to provide the image search method of a kind of integrated classification and global index, its retrieval is the semantic feature that has adopted picture, thereby improved the precision ratio of searching system, and according to two kinds of retrieval modes of classification designator integrated classification and global index of picture to be checked, the problem that recall ratio when having made up classification error is low, and disaggregated model guarantees that most of picture adopts the mode classification retrieval, range of search is dwindled greatly, the fraction picture adopts the mode of index to retrieve, thus the whole retrieval rate that has improved system.
The present invention is achieved by the following technical solutions:
The image search method of a kind of integrated classification and global index comprises the steps:
A) download pictures to be setting up picture library,
B) picture in picture library is classified according to shape, for each classification, pick out representative samples pictures from picture library, form Sample Storehouse.Extract the classification low-level image feature descriptor of all pictures in Sample Storehouse, and utilize support vector machine to train on the low-level image feature descriptor, obtaining the discriminant of each classification, and the discriminant of all classification forms disaggregated model,
C) utilize disaggregated model that all pictures in picture library are classified, with category label and the semantic feature that obtains picture,
D) extract color characteristic and the shape facility of all pictures in picture library, and color characteristic, shape facility and semantic feature be combined into feature database,
E) linked character storehouse and picture library and category label to be forming record sheet,
F) utilize the local sensitivity hash method to set up the index of feature database,
G) reception from user's picture query request, is extracted color characteristic and the shape facility of picture to be checked, and use disaggregated model to treat the inquiry picture and process, with category label and the semantic feature that obtains picture to be checked,
H) judgement picture to be checked category label be greater than or equal,
I) if the category label of picture to be checked greater than, load from feature database according to record sheet with picture to be checked and have the characteristic set of identical category label, then change step k over to,
J) if the category label of picture to be checked equals, inquire about on index according to color characteristic, shape facility and the semantic feature of picture to be checked, obtaining the characteristic set in feature database,
K) color characteristic, shape facility and the semantic feature of characteristic set and picture to be checked are carried out similarity and calculate, and sort according to the similarity value that calculates, obtaining the ranking results corresponding with record sheet,
L) according to ranking results, the picture in the storehouse that Loads Image from record sheet, and loading result is showed the user.
Above-mentioned steps c) comprise substep: the low-level image feature descriptor that extracts Sample Storehouse, adopting card side's kernel algorithm to carry out High Dimensional Mapping to the low-level image feature descriptor processes, to obtain the High Dimensional Mapping vector, adopt support vector machine that the High Dimensional Mapping vector is trained, to obtain the discriminant Wx+b of each classification in Sample Storehouse, wherein w, b are the parameter that the support vector machine training draws, and x is the High Dimensional Mapping vector.
Another object of the present invention is to provide the image indexing system of a kind of integrated classification and global index, its retrieval is the semantic feature that has adopted picture, thereby improved the precision ratio of searching system, and according to two kinds of retrieval modes of classification designator integrated classification and global index of picture to be checked, the problem that recall ratio when having made up classification error is low, and disaggregated model guarantees that most of picture adopts the mode classification retrieval, range of search is dwindled greatly, the fraction picture adopts the mode of index to retrieve, thus the whole retrieval rate that has improved system.
The image indexing system of a kind of integrated classification and global index, comprise: download module, disaggregated model training module, Images Classification module, characteristic extracting module, record sheet are set up module, index module, request processing module, retrieval module, similarity acquisition module, result and are returned to module, download module is used for download pictures to set up picture library, at first the disaggregated model training module classifies according to shape to the picture in picture library, for each classification, pick out representative samples pictures from picture library, form Sample Storehouse.then extract the classification low-level image feature descriptor of all pictures in Sample Storehouse, and utilize support vector machine to train on the low-level image feature descriptor, to obtain the discriminant of each classification, and the discriminant of all classification forms disaggregated model, the Images Classification module is used for utilizing disaggregated model that all pictures of picture library are classified, with category label and the semantic feature that obtains picture, characteristic extracting module is extracted color characteristic and the shape facility of all pictures in picture library, and with color characteristic, shape facility and semantic feature are combined into feature database, record sheet is set up the path of all pictures in module relation feature database and picture library and category label to form record sheet, index module is used for utilizing the local sensitivity hash method to set up the index of feature database, request processing module is used for receiving the picture query request from the user, extract color characteristic and the shape facility of picture to be checked, using disaggregated model to treat the inquiry picture processes, with category label and the semantic feature that obtains picture to be checked, retrieval module be used for judging the category label of picture to be checked be greater than or equal, if the category label of picture to be checked greater than, load entry corresponding in the characteristic set that has an identical category label with picture to be checked and record sheet from feature database according to record sheet, if the category label of picture to be checked equals, according to the color characteristic of picture to be checked, shape facility and semantic feature are inquired about on index, to obtain entry corresponding in characteristic set and the record sheet in feature database, the similarity acquisition module is used for the color characteristic to characteristic set and picture to be checked, shape facility and semantic feature are carried out similarity and are calculated, and sort according to the similarity value that calculates, to obtain the ranking results related with entry, result is returned to module and is used for Load Image picture in the storehouse of picture path according to entry, and loading result is showed the user.
the Images Classification module comprises that the characteristic of division descriptor extracts submodule, High Dimensional Mapping submodule and training submodule, the characteristic of division descriptor extracts the low-level image feature descriptor that submodule is used for extracting Sample Storehouse, the High Dimensional Mapping submodule is used for adopting card side's kernel algorithm to carry out the High Dimensional Mapping processing to the low-level image feature descriptor, to obtain the High Dimensional Mapping vector, the training submodule is used for adopting support vector machine that the High Dimensional Mapping vector is trained, to obtain the discriminant Wx+b of each classification in Sample Storehouse, w wherein, b is the parameter that the support vector machine training draws, x is the High Dimensional Mapping vector.
The present invention has following advantage and technique effect:
1, the inquiry accuracy rate is high
System not exclusively depends on the retrieval character of image, adopts disaggregated model that image library has been carried out once presorting, and the picture in same class has similarity semantically.Utilize disaggregated model to produce the semantic information of picture, make the similarity of machine judgement differentiate more sense organ near the people when retrieval.The retrieval of whole system has incorporated the retrieval effectiveness of disaggregated model, and the query rate of system is improved;
2, fast response time
After image was classified, the amount of images inside each classification greatly reduced, and has accelerated inquiry velocity.Classification can guarantee that the image more than 90% correctly classifies, and residue adopts indexed mode to retrieve on the overall situation less than 10% image.Thereby in the situation that guarantee that inquiry is effective, comprehensively accelerated inquiry velocity;
3, recall ratio is high
The discrimination threshold of disaggregated model is controlled strict, and the image of correct classification retrieves its precision ratio in classification and recall ratio can be guaranteed.For the parts of images that can not correctly classify, adopt global index's mode to retrieve, avoided due to the inaccurate precision ratio that causes of classification and looked into complete low problem.
Description of drawings:
Fig. 1 is the process flow diagram of the image search method of integrated classification of the present invention and global index.
Fig. 2 is the refinement process flow diagram of step in the inventive method (c).
Fig. 3 is the schematic block diagram of the image indexing system of integrated classification of the present invention and global index.
Fig. 4 is the refinement block diagram of Images Classification module in system of the present invention.
Embodiment:
Below at first technical term of the present invention is explained and illustrated:
Representative samples pictures: the feature that can embody a certain classification in picture library.
Semantic feature: picture is when utilizing disaggregated model to classify, and each classification obtains a score value, the vector that the score value of all categories forms.
Category label: picture is when utilizing disaggregated model to classify, and the classification that the maximum score value that obtains is corresponding is numbered.
Color characteristic: the picture color histogram feature of extraction.
Shape facility: the histogram of gradients feature of the pyramid of image.
The local sensitivity hash method: i.e. Local sensitivity Hashing, produce at random one group of vector, according to the vector of random generation, proper vector is distributed to a kind of hash method in different buckets.
Low-level image feature descriptor: the integer vectors that the proper vector of quantized image forms.
The side's of card kernel algorithm: a kind of algorithm of lower dimensional space data-mapping to higher dimensional space.
HOG: i.e. Histogram of Oriented Gradients, gradient orientation histogram can reflect the shape facility of image.
PHOG: i.e. Pyramid Histogram of Oriented Gradients, the HOG feature of pyramid has merged a kind of shape facility of spatial information.
SIFT: i.e. Scale-Invariant Feature Transform, the conversion of yardstick invariant features, a kind of Local Feature Extraction.
GRIDSIFT: i.e. Grid Dense Scale-Invariant Feature Transform, the yardstick invariant features conversion that grid is intensive, the image SIFT feature that adopts the intensive sampling method to obtain.
PGRIDSIFT: i.e. Pyramid Grid Dense Scale-Invariant Feature Transform, the GRIDSIFT feature of pyramid has merged a kind of local feature of spatial information.
The High Dimensional Mapping vector: the lower dimensional space data transformation is to the high-dimensional data of higher dimensional space.
As shown in Figure 1, the image search method of integrated classification of the present invention and global index comprises the following steps:
A) download pictures is to set up picture library;
B) picture in picture library is classified according to shape, for each classification, pick out representative samples pictures from picture library, form Sample Storehouse.Extract the classification low-level image feature descriptor of all pictures in Sample Storehouse, and utilize support vector machine to train on the low-level image feature descriptor, obtaining the discriminant of each classification, and the discriminant of all classification forms disaggregated model;
C) utilize disaggregated model that all pictures in picture library are classified, with category label and the semantic feature that obtains picture;
D) extract color characteristic and the shape facility of all pictures in picture library, and color characteristic, shape facility and semantic feature are combined into feature database, specifically adopt following two kinds of methods;
(d-1) Color Feature Extraction Method: image is showed in rgb space.The pixel of the 24bit of RGB is converted to the value of a 9bit.Method is as follows: each passage has 8 bit positions, at first takes out the highest 3 bit positions of each passage.To R, G, three passages of B have 9 bit, and these 9 bit consist of a numeral, and maximal value is 2
9The number of times that statistics 9bit value occurs, and quantize to form 512 proper vectors of tieing up.
(d-2) method for extracting shape features:
(d-2-1) input picture is converted to gray level image Gray;
(d-2-2) use the canny operator the edge Edge that asks for gray level image Gray;
(d-2-3) to gray level image Gray, ask direction gradient GradientX on level and vertical direction, GradientY asks the comprehensive gradient G radientR of 2 directions;
(d-2-4) according to direction gradient value obtained in the previous step, ask the angle A ngle=atan (GradientY/GradientX) of each pixel, and the angular quantification to 8 of each a pixel interval.Quantization method is
And it is rounded, each is worth in the interval;
(d-2-5) obtain angle matrix M atrixAngle, deposit the quantized value of angle.Gradient matrix MatrixGradient deposits the comprehensive gradient of pixel.For edge image Edge, be 0 point at the edge, MatrixAngle, MatrixGradient are 0; Be not 0 point, MatrixAngle deposits the value after its angular quantification, and MatrixGradient deposits comprehensive Grad;
(d-2-6) ask for the hog feature, statistics has the number that angle after quantification has identical value in image range, and the comprehensive gradient of these pixels is added up, and each angle obtains a feature, has 8 features;
(d-2-7) Pyramidization hog feature obtains the PHOG feature.Image is divided into Isosorbide-5-Nitrae, and the hog feature is asked respectively in 16,64 zonules in each zone, forms a large vector.Have the proper vector of 8 * (1+4+16+64)=680 dimensions;
E) and linked character storehouse and picture library and category label to form record sheet;
F) utilize the local sensitivity hash method to set up the index of feature database;
G) reception from user's picture query request, is extracted color characteristic and the shape facility of picture to be checked, uses disaggregated model to treat the inquiry picture and processes, with category label and the semantic feature that obtains picture to be checked;
H) category label of judgement picture to be checked is greater than 0 or equals 0;
I) if the category label of picture to be checked greater than 0, loads the characteristic set that has the identical category label with picture to be checked from feature database according to record sheet, then change step (k) over to;
J) if the category label of picture to be checked equals 0, inquire about on index according to color characteristic, shape facility and the semantic feature of picture to be checked, to obtain the characteristic set in feature database;
K) color characteristic, shape facility and the semantic feature of characteristic set and picture to be checked are carried out similarity and calculate, and sort according to the similarity value that calculates, to obtain the ranking results corresponding with record sheet;
(k-1) obtain the characteristic distance value: color characteristic adopts the JSD distance
Compare, semantic feature and shape facility adopt Euclidean distance (∑ (x
k-y
k)
2) compare, obtain respectively distance value set { d
k;
(k-2) obtain the similarity value: by presetting the weights set { w of each feature
kAnd distance value set { d obtained in the previous step
k, obtain similarity
Each similarity value is in [0,1] scope;
(k-3) obtain entry, the similarity value of obtaining according to previous step sorts, and loads corresponding entry in record sheet.
L) according to ranking results, the picture in the storehouse that Loads Image from record sheet, and loading result is showed the user.
As shown in Figure 2, in the inventive method, step (c) comprises following substep:
(c1) extract the low-level image feature descriptor of Sample Storehouse;
(c1-1) rasterizing image S
I, image S
IThe blockage of size such as be divided into, extract the SIFT feature of image in blockage, the image SIFT that then this employing rasterizing obtains after processing is called the GRIDSIFT feature.The characteristic set that obtains from all blockages of an image is T={T
k| k=1,2 ... N}, wherein T
k=(D
k, F
k), | N| is image S
IThe number of middle blockage, D
kBe 128 dimensional vectors of the SIFT feature extracted in blockage, F
k=(X
k, Y
k, B
k, H
k) be the descriptor of blockage, (X
k, Y
k) be the center information of blockage, B
kBe the length of side length of blockage, H
kControl threshold value for the regional SIFT feature of blockage;
(c1-2) the image GRIDSIFT proper vector that obtains from step (c1-1) is carried out the pyramid processing, the vector after processing is called PGRIDSIFT.The mode of pyramid is, the length of side that makes blockage value successively is { 4,6,8,10}.Extract the GRIDDSIFT feature of image on the blockage of different size, preserve according to length of side size order.Can intersect through the PGRIDSIFT feature of pyramid and cover whole image, thereby reach the global characteristics of fused images in local feature, become the low-level image feature of Description Image more all sidedly;
(c1-3) extract the PGRIDSIFT feature of sample image according to the method in step (c1-2).Adopt ELKAN KMEANS algorithm to carry out cluster to the PGRIDSIFT feature of all sample images, form K cluster centre, cluster centre number K is through manually being adjusted to value preferably.The cluster centre set is the reference base of being determined by sample.The cluster centre that obtains is preserved into a matrix, and set up the KDTREE index on the cluster centre matrix;
(c1-4) use image S
IPGRIDSIFT feature { T
kCarry out one query on the KDTREE that obtains obtain its subscript index { L in step (c1-3)
k.Blockage positional information { F in the PGRIDSIFT feature of image
k(X
k), F
k(Y
k) at image S
IWidth and height on once quantize, obtain quantized value
According to the row statistics
Form the vectorial S of one dimension statistics.The vectorial S of statistics is quantized into histogram,
Form image S
IThe histogram descriptor H of low-level image feature.
(c2) adopt card side's kernel algorithm to carry out High Dimensional Mapping to the low-level image feature descriptor and process, to obtain the High Dimensional Mapping vector;
(c3) adopt support vector machine that the High Dimensional Mapping vector is trained, to obtain the discriminant Wx+b of each classification in Sample Storehouse, wherein w, b are the parameter that the support vector machine training draws, and x is the High Dimensional Mapping vector.
As shown in Figure 3, the image indexing system of integrated classification of the present invention and global index comprises that module download module 1, disaggregated model training module 2, Images Classification module 3, characteristic extracting module 4, record sheet set up module 5, index module 6, request processing module 7, retrieval module 8, similarity acquisition module 9, result and return to module 10.
Download module 1 is used for download pictures to set up picture library.
At first disaggregated model training module 2 classifies according to shape to the picture in picture library, for each classification, picks out representative samples pictures from picture library, forms Sample Storehouse.Then extract the classification low-level image feature descriptor of all pictures in Sample Storehouse, and utilize support vector machine to train on the low-level image feature descriptor, obtaining the discriminant of each classification, and the discriminant of all classification forms disaggregated model.
Images Classification module 3 is used for utilizing disaggregated model that all pictures of picture library are classified, with category label and the semantic feature that obtains picture.
Characteristic extracting module 4 is extracted color characteristic and the shape facility of all pictures in picture library, and color characteristic, shape facility and semantic feature are combined into feature database.
Record sheet is set up the path of all pictures in module 5 linked character storehouses and picture library and category label to form record sheet.
Index module 6 is used for utilizing the local sensitivity hash method to set up the index of feature database.
Request processing module 7 is used for receiving the picture query request from the user, extracts color characteristic and the shape facility of picture to be checked, uses disaggregated model to treat the inquiry picture and processes, with category label and the semantic feature that obtains picture to be checked.
Retrieval module 8 is used for judging that the category label of picture to be checked is greater than 0 or equals 0, if the category label of picture to be checked is greater than 0, load entry corresponding in the characteristic set that has an identical category label with picture to be checked and record sheet from feature database according to record sheet, if the category label of picture to be checked equals 0, inquire about on index according to color characteristic, shape facility and the semantic feature of picture to be checked, to obtain entry corresponding in characteristic set and the record sheet in feature database.
Similarity acquisition module 9 is used for that color characteristic, shape facility and the semantic feature of characteristic set and picture to be checked are carried out similarity and calculates, and sorts according to the similarity value that calculates, to obtain the ranking results related with entry.
Result is returned to module 10 and is used for picture path according to the entry picture in storehouse that Loads Image, and loading result is showed the user.
As shown in Figure 4, module comprises that the characteristic of division descriptor extracts submodule 31, High Dimensional Mapping submodule 32, High Dimensional Mapping submodule 32, training submodule 33.
The characteristic of division descriptor extracts the low-level image feature descriptor that submodule 31 is used for extracting Sample Storehouse;
High Dimensional Mapping submodule 32 is used for adopting card side's kernel algorithm to carry out the High Dimensional Mapping processing to the low-level image feature descriptor, to obtain the High Dimensional Mapping vector;
Training submodule 33 is used for adopting support vector machine that the High Dimensional Mapping vector is trained, and to obtain the discriminant of each classification in Sample Storehouse, wherein w, b are the parameter that the support vector machine training draws, and x is the High Dimensional Mapping vector.
Example
Test data source: 02 large class in Chinese design patent, i.e. clothes and haberdashery, totally 63350 records, 12 large classes, i.e. transportation or hoisting tool, totally 65416 records.System's summary journal number 128766.
Test environment of the present invention is as shown in table 1:
| CPU | Internal memory | Hard disk | Operating system | Amount of images |
| Intel Core(TM)i7 | 6G | 1T | X86_64GNU/LINUX | 128766 |
Table 1 test environment
According to test result, the response speed that system is inquired about on other data of 10w level is in 1s.Average inquiry accuracy rate is more than 85%, and recall level average is more than 80%.
Response speed is that the user submits to picture to the time that returns results, and does not comprise network latency.The inquiry accuracy rate is that all that return return results the shared ratio of middle similar pictures.Recall ratio is the ratio that similar pictures that system returns accounts for all similar pictures in picture library.
Claims (2)
1. the image search method of an integrated classification and global index, is characterized in that, comprises the steps:
A) download pictures is to set up picture library;
B) picture in described picture library is classified according to shape, for each classification, pick out representative samples pictures from described picture library, form Sample Storehouse, extract the classification low-level image feature descriptor of all pictures in described Sample Storehouse, and utilize support vector machine to train on described low-level image feature descriptor, obtaining the discriminant of each classification, and the discriminant of all classification forms disaggregated model;
C) utilize described disaggregated model that all pictures in described picture library are classified, with category label and the semantic feature that obtains described picture; Described step (c) comprises following substep:
(c1) extract the low-level image feature descriptor of described Sample Storehouse;
(c2) adopt card side's kernel algorithm to carry out High Dimensional Mapping to described low-level image feature descriptor and process, to obtain the High Dimensional Mapping vector;
(c3) adopt support vector machine that described High Dimensional Mapping vector is trained, to obtain the discriminant Wx+b of each classification in described Sample Storehouse, wherein W, b are the parameter that the support vector machine training draws, and x is described High Dimensional Mapping vector;
D) extract color characteristic and the shape facility of all pictures in described picture library, and described color characteristic, described shape facility and described semantic feature are combined into feature database;
E) related described feature database and described picture library and described category label are to form record sheet;
F) utilize the local sensitivity hash method to set up the index of described feature database;
G) reception from user's picture query request, is extracted color characteristic and the shape facility of picture to be checked, uses described disaggregated model that described picture to be checked is processed, with category label and the semantic feature that obtains described picture to be checked;
H) category label of the described picture to be checked of judgement is greater than 0 or equals 0;
I) if the category label of described picture to be checked greater than 0, loads the characteristic set that has the identical category label with described picture to be checked from described feature database according to described record sheet, then change step (k) over to;
J) if the category label of described picture to be checked equals 0, inquire about on described index according to color characteristic, shape facility and the semantic feature of described picture to be checked, to obtain the characteristic set in described feature database;
K) color characteristic, shape facility and the semantic feature of described characteristic set and described picture to be checked are carried out similarity and calculate, and sort according to the similarity value that calculates, to obtain the ranking results corresponding with described record sheet;
L) according to described ranking results, load the picture in described picture library from described record sheet, and loading result is showed the user.
2. the image indexing system of an integrated classification and global index, comprise: download module (1), disaggregated model training module (2), Images Classification module (3), characteristic extracting module (4), record sheet are set up module (5), index module (6), request processing module (7), retrieval module (8), similarity acquisition module (9), result and are returned to module (10), it is characterized in that
Described download module (1) is used for download pictures to set up picture library;
Described disaggregated model training module (2) is at first classified according to shape to the picture in described picture library, for each classification, pick out representative samples pictures from described picture library, form Sample Storehouse, then extract the classification low-level image feature descriptor of all pictures in described Sample Storehouse, and utilize support vector machine to train on described low-level image feature descriptor, obtaining the discriminant of each classification, and the discriminant of all classification forms disaggregated model;
Described Images Classification module (3) is used for utilizing described disaggregated model that all pictures of described picture library are classified, with category label and the semantic feature that obtains described picture;
Described Images Classification module (3) comprises that the characteristic of division descriptor extracts submodule (31), High Dimensional Mapping submodule (32) and training submodule (33);
Described characteristic of division descriptor extracts the low-level image feature descriptor that submodule (31) is used for extracting described Sample Storehouse;
Described High Dimensional Mapping submodule (32) is used for adopting card side's kernel algorithm to carry out the High Dimensional Mapping processing to described low-level image feature descriptor, to obtain the High Dimensional Mapping vector;
Described training submodule (33) is used for adopting support vector machine that described High Dimensional Mapping vector is trained, to obtain the discriminant Wx+b of each classification in described Sample Storehouse, wherein W, b are the parameter that the support vector machine training draws, and x is described High Dimensional Mapping vector;
Described characteristic extracting module (4) is extracted color characteristic and the shape facility of all pictures in described picture library, and described color characteristic, described shape facility and described semantic feature are combined into feature database;
Described record sheet is set up the path of all pictures in the related described feature database of module (5) and described picture library and described category label to form record sheet;
Described index module (6) is used for utilizing the local sensitivity hash method to set up the index of described feature database;
Described request processing module (7) is used for receiving the picture query request from the user, extract color characteristic and the shape facility of picture to be checked, use described disaggregated model that described picture to be checked is processed, with category label and the semantic feature that obtains described picture to be checked;
described retrieval module (8) is used for judging that the category label of described picture to be checked is greater than 0 or equals 0, if the category label of described picture to be checked is greater than 0, load entry corresponding in the characteristic set that has an identical category label with described picture to be checked and described record sheet from described feature database according to described record sheet, if the category label of described picture to be checked equals 0, according to the color characteristic of described picture to be checked, shape facility and semantic feature are inquired about on described index, to obtain entry corresponding in characteristic set and the described record sheet in described feature database,
Described similarity acquisition module (9) is used for that color characteristic, shape facility and the semantic feature of described characteristic set and described picture to be checked are carried out similarity and calculates, and sort according to the similarity value that calculates, to obtain the ranking results related with described entry;
Described result is returned to module (10) and is used for according to the picture in the described picture library of picture path loading of described entry, and loading result is showed the user.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN 201110423143 CN102521366B (en) | 2011-12-16 | 2011-12-16 | Image retrieval method integrating classification with hash partitioning and image retrieval system utilizing same |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN 201110423143 CN102521366B (en) | 2011-12-16 | 2011-12-16 | Image retrieval method integrating classification with hash partitioning and image retrieval system utilizing same |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN102521366A CN102521366A (en) | 2012-06-27 |
| CN102521366B true CN102521366B (en) | 2013-06-26 |
Family
ID=46292279
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN 201110423143 Expired - Fee Related CN102521366B (en) | 2011-12-16 | 2011-12-16 | Image retrieval method integrating classification with hash partitioning and image retrieval system utilizing same |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN102521366B (en) |
Families Citing this family (22)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103336795B (en) * | 2013-06-09 | 2017-03-08 | 华中科技大学 | Video index method based on multiple features |
| CN103744964A (en) * | 2014-01-06 | 2014-04-23 | 同济大学 | Webpage classification method based on locality sensitive Hash function |
| CN105224582B (en) * | 2014-07-03 | 2018-11-09 | 联想(北京)有限公司 | Information processing method and equipment |
| CN105446997B (en) * | 2014-07-22 | 2019-06-18 | 华为技术有限公司 | Image content recognizing method and device |
| CN104199922B (en) * | 2014-09-01 | 2019-05-03 | 中国科学院自动化研究所 | A kind of large-scale image library searching method based on the similar hash algorithm in part |
| CN104463922B (en) * | 2014-12-03 | 2017-09-08 | 天津大学 | A kind of characteristics of image coding and recognition methods based on integrated study |
| CN106557523B (en) * | 2015-09-30 | 2020-05-12 | 佳能株式会社 | Representative image selection method and apparatus, and object image retrieval method and apparatus |
| CN106203493B (en) * | 2016-07-04 | 2019-07-12 | 何广森 | A kind of food identification device and recognition methods |
| CN106844421A (en) * | 2016-11-30 | 2017-06-13 | 上海仙剑文化传媒股份有限公司 | A kind of digital picture management method and system |
| CN106844791A (en) * | 2017-03-28 | 2017-06-13 | 马佳 | A kind of search method of feature based cluster |
| CN107180079B (en) * | 2017-04-28 | 2020-04-28 | 华南理工大学 | Image retrieval method based on convolutional neural network and tree and hash combined index |
| CN107577990B (en) * | 2017-08-09 | 2020-02-18 | 武汉世纪金桥安全技术有限公司 | Large-scale face recognition method based on GPU (graphics processing Unit) accelerated retrieval |
| CN109241319B (en) * | 2018-09-28 | 2021-11-02 | 百度在线网络技术（北京）有限公司 | Picture retrieval method, device, server and storage medium |
| CN109242042B (en) * | 2018-09-28 | 2019-12-20 | 百度在线网络技术（北京）有限公司 | Picture training sample mining method and device, terminal and computer readable storage medium |
| CN109947976B (en) * | 2019-01-22 | 2023-05-16 | 湖南城市学院 | System and method for analyzing Chinese traditional element embodiment in modern art design |
| CN110084642A (en) * | 2019-04-19 | 2019-08-02 | 哈尔滨新中新电子股份有限公司 | Vegetable recognition methods based on target detection and image retrieval |
| CN110175249A (en) * | 2019-05-31 | 2019-08-27 | 中科软科技股份有限公司 | A kind of search method and system of similar pictures |
| CN110413813B (en) * | 2019-06-25 | 2023-05-12 | 宁波图达信息技术有限公司 | Method for searching same or similar images |
| CN110413647B (en) * | 2019-07-08 | 2021-04-06 | 上海鸿翼软件技术股份有限公司 | High-dimensional vector unequal length sequence similarity rapid calculation system |
| CN111930985A (en) * | 2020-07-08 | 2020-11-13 | 泰康保险集团股份有限公司 | Image retrieval method and device, electronic equipment and readable storage medium |
| CN114049508B (en) * | 2022-01-12 | 2022-04-01 | 成都无糖信息技术有限公司 | Fraud website identification method and system based on picture clustering and manual research and judgment |
| CN117392440B (en) * | 2023-09-26 | 2024-03-01 | 武汉纺织大学 | Textile fabric retrieval method and system based on tissue structure and color classification |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7171042B2 (en) * | 2000-12-04 | 2007-01-30 | Intel Corporation | System and method for classification of images and videos |
| CN101710334B (en) * | 2009-12-04 | 2012-01-25 | 大连理工大学 | Large-scale image library retrieving method based on image Hash |
-
2011
- 2011-12-16 CN CN 201110423143 patent/CN102521366B/en not_active Expired - Fee Related
Also Published As
| Publication number | Publication date |
|---|---|
| CN102521366A (en) | 2012-06-27 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN102521366B (en) | Image retrieval method integrating classification with hash partitioning and image retrieval system utilizing same | |
| Zheng et al. | SIFT meets CNN: A decade survey of instance retrieval | |
| Bai et al. | VHR object detection based on structural feature extraction and query expansion | |
| Jang et al. | Car-Rec: A real time car recognition system | |
| CN110059807A (en) | Image processing method, device and storage medium | |
| CN102542058A (en) | Hierarchical landmark identification method integrating global visual characteristics and local visual characteristics | |
| CN105574063A (en) | Image retrieval method based on visual saliency | |
| CN105740378B (en) | Digital pathology full-section image retrieval method | |
| Vieux et al. | Content based image retrieval using bag-of-regions | |
| CN104317946A (en) | Multi-key image-based image content retrieval method | |
| CN111414958B (en) | Multi-feature image classification method and system for visual word bag pyramid | |
| CN108319959A (en) | A kind of corps diseases image-recognizing method compressed based on characteristics of image with retrieval | |
| Úbeda et al. | Improving pattern spotting in historical documents using feature pyramid networks | |
| CN112182264B (en) | Method, device and equipment for determining landmark information and readable storage medium | |
| JP6017277B2 (en) | Program, apparatus and method for calculating similarity between contents represented by set of feature vectors | |
| CN103049570A (en) | Method for searching and sorting images and videos on basis of relevancy preserving mapping and classifier | |
| Ghosh et al. | Efficient indexing for query by string text retrieval | |
| Bai et al. | Informative patches sampling for image classification by utilizing bottom-up and top-down information | |
| Tsai | Two strategies for bag-of-visual words feature extraction | |
| Kumar et al. | Automatic feature weight determination using indexing and pseudo-relevance feedback for multi-feature content-based image retrieval | |
| Tencer et al. | Sketch-based retrieval of document illustrations and regions of interest | |
| Selvam et al. | A new architecture for image retrieval optimization with HARP algorithm | |
| Gao et al. | An Interactive Appearance-based Document Retrieval System for Historical Newspapers. | |
| Kiruthika et al. | A Comparative Study Of Image Retrieval Techniques | |
| Radha et al. | Modified cuckoo search algorithm: feature subset selection & shape, color and texture features descriptors for content-based image retrieval |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20130626 Termination date: 20151216 |
|
| EXPY | Termination of patent right or utility model |